Fig. 2: Task design and the procedure of decoding analyses.
A Sequence of trial events in the rule-selection task with the response deadline. In the variable or sampled SOA phase, an audio trigger signal indicated the start of a response deadline time window. B Spatial translation of different rules (rows) mapping different stimuli (columns) to responses (arrows), yielding 12 independent conjunctions. C Schematic of the time-resolved representational similarity analysis. For each sample time (t), a scalp-distributed pattern of EEG was used to decode the specific rule/stimulus/response configuration of a required action. The decoder produced sets of classification probabilities for each of the possible action constellations. The profile of classification probabilities reflects the similarity structure of the underlying representations, where action constellations with shared features are more likely to be confused. For each trial and timepoint, the classification probabilities were regressed onto model vectors as predictors that reflect the different, possible representations. In each model matrix, the shading of squares indicates the theoretically predicted classification probabilities (darker shading means higher probabilities) in all possible pairs of constellations. The coefficients associated with each predictor (i.e., t-values) reflect the unique variance explained by each of the constituent features and their conjunction. D Schematic of the time-resolved binary classification method used to estimate the representational dimensionality51. For each time point (t), a pattern of EEG associated with unique action constellations or input conditions (c1–12) takes a position in the multidimensional neural space spanning in r1-e dimensions. By assigning new binary class labels (e.g., A or B) to the input conditions, we can generate arbitrary binary groupings given the task conditions. Because the higher dimensional geometry of neural responses generally affords more arbitrary linear separations by being more expressive, the count of successfully implementable binary classifications of newly defined groupings scales with the representational dimensionality. To adapt the method to EEG by lowering the cutoff threshold for classification, we used an exclusive cutoff method. In this method, classifications must exceed the cutoff threshold in all the different input conditions assigned to each trial, as opposed to the averaged performance across c1–12, to be marked as successful.
